Running support/MivoChat field guide

Reduce AI chatbot hallucinations on product docs

A support bot is only useful when answers stay within the product documentation and unknowns reach a person.

By MivoChat product team · Reduce AI chatbot hallucinations on product docs · Published and updated October 12, 2026

Quick answer

Keep public docs current, connect the right pages, test both answerable and unsupported questions, and require a clear handoff when sources do not support a claim.

Installation steps

    Make the source set dependable

    Remove stale versions, resolve conflicting pricing statements, and ensure docs pages name the product version they describe. Include public API reference and troubleshooting pages that match what customers use. Compare the crawler source list with the published documentation index and remove duplicate copies that disagree on defaults or limits. Give each major feature one canonical page with a visible version and last-updated date, then link migration notes to the current instructions instead of leaving both versions equally prominent.

    Test grounded and ungrounded prompts

    Try “How do I rotate my webhook secret?”, an edge case with a version constraint, and a question absent from the docs. Check that answers stay grounded in relevant pages and the unsupported question is not filled with a plausible guess. Keep a regression sheet with each prompt, expected page, required caveat, and acceptable handoff result; rerun it after documentation releases or a change to the connected source set. A response that sounds correct but points to the wrong version should fail the check.

    Treat handoff as part of the answer

    Give the visitor a route to Slack or Telegram support when the docs are incomplete. Review handoffs regularly and fix source gaps; do not turn an unverified support reply into a published answer automatically. When a visitor reports a current outage or private account state, treat the chat as a route to a teammate rather than a source of operational status. Assign one product owner to approve new policy and limit changes before those statements enter public docs.

    Before you change the setup

    current API and product pages, release ownership, examples of incorrect answers, and an available human handoff

    the canonical docs, source crawl scope, and a repeatable set of answerable and unanswerable test prompts

    better sources and tests reduce unsupported answers; they do not guarantee every model response is correct

    Test the support boundary

    • What exact scopes does the token endpoint require?
    • How do I rotate a webhook secret with zero downtime?
    • What is our policy if a behavior is not documented?

    Frequently asked questions

    Conflicting release pages produce different answers

    Label versioned instructions clearly, retire obsolete pages from the crawl, and make the canonical version easy to identify.

    The model invents a missing default value

    Add the value and its version boundary to the source, then add the unsupported prompt to the regression set.

    A broad page wins over the exact API reference

    Improve the exact page title and navigation context, then test whether the answer points to that endpoint instead of a general overview.

    The unknown question gets no clear next step

    Make the handoff route visible and verify the connected teammate channel receives a new conversation before launch.

    Keep exploring